design-ai-product-evaluation

Pass

Audited by Gen Agent Trust Hub on Oct 27, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDATA_EXFILTRATION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to ingest external organizational data to build evaluation tasks and rubrics, creating a potential surface for instructions embedded in that data to influence agent behavior.
  • Ingestion points: SKILL.md identifies signals, surveys, projects, and documents as primary sources for tasks.
  • Boundary markers: The skill includes a high-level instruction to "govern sensitive data" and "keep synthetic judgments distinct," but lacks specific delimiters or warnings to ignore embedded instructions in the ingested sources.
  • Capability inventory: The skill relies on the BuildBetter MCP to search and retrieve organizational content, which the agent then processes to generate rubrics and datasets.
  • Sanitization: No specific sanitization, escaping, or filtering of the retrieved content is defined within the instructions.
  • [DATA_EXPOSURE]: The skill encourages the retrieval of broad organizational data into the agent's context, which may include sensitive information.
  • Evidence: The workflow in SKILL.md suggests using customer outcomes, signals, and internal documents via MCP calls to support evaluation design.
Audit Metadata
Risk Level
SAFE
Analyzed
Oct 27, 2026, 03:14 AM